Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

stDyer-image improves clustering analysis of spatially resolved transcriptomics and proteomics with morphological images.

Bioinformatics · 2026
L1 75/100 3/4
Why this verdict

The main results reproduced: recomputed values matched the published ones within tolerance.

Reproduced on the brainbox compute brainarbeit.com
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Concordant (toward reproduced)
All content-critical questions reproduced
-4 pts
From: Q7 · Core claim 🟢
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Every question reproduced
-1 pts
From: “every question reproduced”
Total score -7
✓ What held up
  • Same input data as the authors
  • Reported values were directly comparable
  • No relevant deviation in data/preprocessing
  • No authors-side cause for any deviation
  • Reported values are derivable from the shared data
  • Any deviation was negligible
  • The central claim held under reproduction
  • Overall, the reproduction was clean
What did not (or only partly)
  • Every checked point held up.
How its reproducibility compares
75/100
Reproducibility score
at the mean
vs. all fields · 1173 studies
🎯 Scores higher than 45% of all assessed papers rank 612 of 1173 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

DESCRIBED WELL ENOUGH: yes. stDyer-image (authors' own repo @e2ae94e; Zenodo 15243986) is reproducible. The headline Mouse_brain (Stereo-seq) ARI 0.517 (Fig 3c) is reproduced EXACTLY (0.5169) from the authors' DEPOSITED pipeline output (checkpoints/Mouse_brain/Mouse_brain.h5ad obs.pred_labels) using the paper's exact metric adjusted_rand_score over labelled units -> the headline number is fully data-backed, no fabrication indication; gt_labels-vs-annotation sanity = 1.0. This is 1:1 on the deposited result. We additionally built a functional env from the pinned stdyer_image.yml (dropping the unused rapids/cuml stack) and confirmed the shipped code RUNS END-TO-END on the paper's data into the training loop; the full from-scratch multi-epoch retrain (the stronger code-generates-the-number test) was NOT completed because the faithful run is GPU-only and all 4 H100 nodes were saturated (~5h queue) while the CPU fallback's per-epoch image-kNN is too slow for 200 epochs -> deliberately stopped per the 80/20 rule. NOT ATTEMPTED: the other reported ARIs (NSCLC CosMx 0.561 / CODEX 0.400/0.361/0.346, all multi-slice), secondary SI/FMI panels, SVG panels, and the 8 competitor tools. Full per-unit reproduced labels saved locally (reproduction/outputs/repro_pred_labels.csv.gz, 38736 rows) so a human can recompute the ARI by hand.

💻 Code ↗ 🗄 Data: 10.5281/zenodo.15243986

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

Assessment versions

Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.

  1. v1 current initial assessment Score 75
    assessed: 2026-06-16 ⛓ 7a3f739a071e
✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-06-16
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-16
no human curator yet
Last updated
2026-08-05

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: opus
Founding hypothesis

The authors hypothesize that morphological images contain sufficient information to directly infer cluster labels (as a pathologist would), so directly linking image features to cluster labels—rather than using images to enhance gene expression data—can improve clustering of spatially resolved transcriptomics and proteomics datasets.

Core claims
  • stDyer-image is an end-to-end deep learning framework that directly associates image features with predicted cluster labels to improve clustering of SRT and SRP data with images. method
  • stDyer-image achieves superior clustering performance compared to eight state-of-the-art tools across five technologies. finding
  • An image-related loss within a GMVAE encourages units with similar image embeddings to share similar soft cluster labels, optimizing predicted labels directly. mechanism
  • stDyer-image is scalable to large datasets via mini-batch neighbor sampling and multi-GPU training, and is compatible with CosMx, Stereo-seq, 10x Xenium, and CODEX. method
  • Unlike methods that enhance gene expression with images, directly linking images to labels avoids introducing noise from genes not reflected by tissue color or shape. finding
  • stDyer-image can identify spatially variable genes using integrated gradient analysis. method
  • Open-source software and tutorials are provided at https://github.com/ericcombiolab/stDyer-image. resource
Experimental setups
Assay System Perturbation Readout Platform
spatially resolved transcriptomics (SRT) with morphology image human non-small-cell lung cancer (NSCLC) tissue, 20 slices, 87606 units, 8 cell types none clustering/cell type and tumor region identification (ARI, SI, FMI) CosMx
spatially resolved transcriptomics (SRT) with morphology image human breast cancer tissue none clustering of spatial domains / cell types 10x Xenium
spatially resolved transcriptomics (SRT) with morphology image mouse brain tissue none clustering of spatial domains Stereo-seq
spatially resolved proteomics (SRP) with morphology image human intestine tissue none clustering / cell type identification CODEX
image feature extraction 250 × 250-pixel image patches from highest-resolution tissue images none image feature embeddings, K-nearest image neighbors by cosine distance ResNet18 pre-trained on morphological image datasets (Ciga et al. 2022)
Key results
  • stDyer-image performed best on the NSCLC dataset across all 20 slices in average ARI. ARI = 0.561
  • stDyer-image obtained the highest Silhouette score on embedding compared to other methods. SI = 0.323
  • stDyer-image obtained the highest Silhouette score on gene expression. SI = -0.019
  • stDyer-image obtained the highest Fowlkes-Mallows index compared to other methods. FMI = 0.717
  • stDyer-image showed superior performance in predicting tumor regions, while stDyer, BayesSpace, and CellCharter tended to split regions.
Key statistics
  • other ARI = 0.561 (average across 20 slices) (stDyer-image average ARI on NSCLC CosMx dataset)
  • other SI = 0.323 (highest Silhouette score on embedding, NSCLC dataset)
  • other SI = -0.019 (highest Silhouette score on gene expression, NSCLC dataset)
  • other FMI = 0.717 (highest Fowlkes-Mallows index, NSCLC dataset)
  • count 87606 units (total units in NSCLC CosMx dataset (20 slices))
  • count 8 cell types (annotations in NSCLC dataset)
  • count top 3000 highly variable genes (gene selection during preprocessing)
  • count 250 × 250 pixels (image patch size used to represent local cell environment)

Statistical methods review

Model: sonnet

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

stDyer-image is a deep-learning benchmarking paper that introduces a GMVAE+GAT clustering framework for spatially resolved transcriptomics/proteomics with morphological images. Performance evaluation is based entirely on unsupervised clustering quality metrics (ARI, Silhouette Index, Fowlkes-Mallows Index) comparing nine methods across multiple datasets and technologies. No classical null-hypothesis significance testing is reported; results are presented as metric scores with box plots and tables. The provided text covers the NSCLC/CosMx dataset (87 606 units, 20 slices) in detail; additional datasets are referenced but the text is truncated.

Replicationunclear Sample sizeDataset size stated (87 606 units, 20 slices for NSCLC/CosMx); no statistical power calculation or sample-size justification described GroupsNine clustering methods (stDyer-image, stDyer, BayesSpace, CellCharter, SpaGCN, stLearn, SiGra, DeepST, MUSE) across five SRT/SRP technologies Pairingna Randomization/blindingnot stated Dispersionunclear Exact p-valuesno Effect sizesno Confidence intervalsno Multiplicity correctionnone stated
Statistical tests used
Test Applied to n Assumptions
Adjusted Rand Index (ARI) Primary clustering evaluation metric across all datasets; single-slice scores shown in box plot (Fig. 2c) and averaged across 20 NSCLC slices (Fig. 2d) 87 606 units across 20 slices (NSCLC/CosMx); per-slice n not stated na
Silhouette Index (SI) Secondary clustering quality metric on gene expression and embedding spaces (Table S1) na
Fowlkes-Mallows Index (FMI) Secondary clustering performance metric (Table S1) na
Approaches that could also have been used
  • Method comparisons are based on point-estimate average ARI across slices with no statistical significance testing
    Could also: A non-parametric pairwise test (e.g. Wilcoxon signed-rank on per-slice ARI scores across the 20 NSCLC slices, with FDR correction for nine pairwise comparisons) could also be used — When per-slice ARI distributions overlap, a formal test quantifies whether observed differences in means exceed what would be expected by chance; this is particularly informative when sample counts (slices or datasets) are moderate
  • ARI alone was used as the primary evaluation metric
    Could also: Normalized Mutual Information (NMI) or the V-measure could also serve as primary or co-primary metrics — Different metrics emphasize different aspects of clustering agreement (NMI is symmetric and bounded [0,1]; V-measure decomposes into homogeneity and completeness); reporting multiple well-defined metrics makes comparisons more robust to metric-specific quirks
  • Box plots of per-slice ARI are shown but the dispersion measure is not explicitly named in the text
    Could also: Explicitly reporting median, IQR, and range, or mean ± SD, alongside box plots would also convey variability — Named dispersion statistics allow numerical comparison across papers and enable readers to assess variability without extracting values from figures
  • Each method's ARI is presented as a single average over all 20 slices without confidence intervals
    Could also: Bootstrap confidence intervals (e.g. 95% CI via percentile bootstrap over the 20 slices) could also be computed for each method's mean ARI — CIs communicate estimation uncertainty and can indicate whether performance differences between methods are practically meaningful given within-dataset variability
  • The clustering number was fixed to the ground-truth number of clusters for all methods during evaluation
    Could also: An additional evaluation using unsupervised cluster-number selection (e.g. via silhouette maximisation or the elbow method) could also be reported — Fixing the cluster count to ground truth favours methods with strong prior knowledge; evaluating under automatic cluster-number selection reflects realistic deployment conditions where ground truth is unknown
  • Benchmarking used a single random initialisation per method per dataset (no mention of multiple runs)
    Could also: Repeating each method's run with multiple random seeds and reporting mean ± SD ARI across seeds could also be done — Deep learning methods with stochastic initialisation can have run-to-run variability; multiple-seed evaluation separates method quality from lucky or unlucky initialisation
Software: Scanpy · Python/NumPy · Spateo · scikit-image (skimage) · ResNet18 (pretrained on morphological images, Ciga et al. 2022)

Citation network

Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.

Authors · 3
Citations
0
Impact: low
Foundation confidence
None of its references are in our reproducibility record yet — its foundation cannot be assessed.
Topics

No assessed neighbours yet — the network grows as more papers are assessed.

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

Scope — pmid-41692960 (stDyer-image)

Paper: stDyer-image: end-to-end deep-learning clustering for spatially resolved transcriptomics (SRT) and proteomics (SRP) with morphology images. Xu K, Zhou XM, Zhang L. Bioinformatics 2026. PMID 41692960 / PMC12960910.

Code: https://github.com/ericcombiolab/stDyer-image (pinned commit e2ae94e919247a41fea817ff793142e0a717d1f2, 2025-01-23 — authors' OWN repo, PyTorch-Lightning + Hydra) Data: Zenodo 10.5281/zenodo.15243986 — single file stDyer-image_analysis.zip (5.9 GB, contains processed analysis datasets)

Reproduction metric

The repo computes the exact paper metric in src/models/gmvgat_model.py:1754: exp_ARI = adjusted_rand_score(annotation_int, pred_labels) over labelled units, prints Epoch N ARI: <val> each epoch and logs test/highest_exp_ARI. The paper (Section 2.5) uses the Adjusted Rand Index (Hubert & Arabie 1985) as the primary clustering metric. So the reproduced number is directly comparable to the reported ARI.

In scope (pipeline-derived; attempted)

  • C1 — Mouse_brain (Stereo-seq/MOSTA), ARI 0.517 (Fig 3c). PRIMARY TARGET. Maps 1:1 onto the shipped configs/experiment/example.yaml (MOSTA Mouse_brain, num_classes=20 inferred from annotation, num_hvg=3000, 200 epochs, single H100 GPU). python run.py experiment=example.yaml → ARI printed/logged. Lowest-hanging fruit.

In scope but deferred (80/20 — larger / multi-slice; attempt only if time)

  • C2 NSCLC CosMx ARI 0.561 (avg of 20 slices) — needs 20 runs + the CosMx data.
  • C3/C4/C5 CODEX colon/small-bowel ARI 0.400/0.361/0.346 — multiple slices, separate configs.

Out of scope (not pipeline / not attempted)

  • Silhouette (SI) and Fowlkes-Mallows (FMI) secondary metrics, qualitative figure panels (visualizations), spatially-variable-gene panels — derivative/illustrative.
  • Benchmark competitor tools (BayesSpace, CellCharter, SpaGCN, stLearn, SiGra, DeepST, MUSE, stDyer): re-running 8 other tools is out of the 80/20 first pass.

Notes

  • GPU deep-learning method (CUDA 11.8, dgl cu117, pytorch, rapids). Stochastic: seed=42 set, but exact 1:1 to 3 decimals not guaranteed across GPU/cuDNN/lib versions — within-tolerance agreement (|Δ| small) is the realistic bar.
  • Env: faithful build from shipped fully-pinned stdyer_image.yml.
C1
Reported
Mouse_brain (Stereo-seq/MOSTA) ARI = 0.517 (stDyer-image, Fig 3c)
Reproduced
ARI = 0.5169 (NMI 0.6389, FMI 0.5567; 20 clusters; n=38736) recomputed from the authors' deposited output pred_labels vs annotation, using the paper's exact metric
exact
C1b
Reported
Mouse_brain ARI 0.517 reproduced by re-running the shipped tool from scratch
Reproduced
Pipeline confirmed RUNNABLE end-to-end on the paper's data (load+HVG+ResNet18 image embeds+graph+model+trainer init+entered training); full multi-epoch retrain not completed (GPU queue ~5h; CPU fallback too slow per 80/20)
partial

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

🤖 AI curator · claude (ai-curator room) · v1.0 L1 75/100

An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.

🟢1. Data identity
🟢2. Endpoint comparability
🟢3. Location of the main deviation
🟢4. Cause of the deviation
🟢5. Derivability / plausibility
🟢6. Severity of the deviation
🟢7. Core claim
🟢8. Severity of the miss (overall human judgment)
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Concordant (toward reproduced)
All content-critical questions reproduced
-4 pts
From: Q7 · Core claim 🟢
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Every question reproduced
-1 pts
From: “every question reproduced”
Total score -7

The headline claim C1 (Mouse_brain Stereo-seq ARI = 0.517, Fig 3c) reproduces exactly (0.5169) by recomputing the paper's metric over the authors' deposited pred_labels, with a gt_labels-vs-annotation sanity ARI of 1.0 — fully data-backed, no fabrication indication. The sole nuance, on our side, is that the exact match comes from the deposited pipeline output rather than a completed from-scratch H100 retrain (confirmed runnable end-to-end but stopped per 80/20 on GPU-queue grounds). Deviation is negligible (rounding only) and the central number holds; the other ARIs (C2-C5) and 8 competitors were out of scope, not contradicted. Overall a clean 1:1 reproduction.

🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.

Reproduction footprint

claude-opus-4-8

Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.

382.9 k
tokens (I/O) · 58.4 M incl. cache
53 min
runtime · 1.27 CPU-h
20.6 GB
peak RAM
15 (2 failed)
HPC jobs
hummel
machine